FDA could confidentially test AI models on their predictions of clinical trial outcomes, then approve drugs without trials after a series of accurate predictions, says Retro CEO Joe Betts-LaCroix

FDA could confidentially test AI models on their predictions of clinical trial outcomes, then approve drugs without trials after a series of accurate predictions, says Retro CEO Joe Betts-LaCroix

At the ARDD conference on October 2, the CEO of Retro Biosciences, which develops therapies for aging, outlined a proposal: FDA would compare AI predictions with trial outcomes, while the model developer would be kept unaware of which trials were being used to test it. A recording of the talk was released on October 26.

According to Betts-LaCroix, AI proposes vast numbers of drug candidates, while around 50 new drugs are approved each year. He sees clinical trials as the bottleneck and wants AI to predict their outcomes. A model that reads the internet learns about a trial after it has ended and reports have been published, so its prediction is partly artificial, he said. The model therefore needs to be tested on trials whose outcomes it does not know.

FDA holds information on trials and their outcomes, he said: companies must tell the agency what they plan to test to obtain permission, and they submit results from the previous phase before proceeding to the next. Less information reaches the public: when FDA published letters explaining its reasons for denying approval in July 2025, it cited a 2015 analysis showing that companies omitted 85% of the agency’s safety and efficacy concerns when announcing those decisions.

Betts-LaCroix sees two ways to use these data. The first is to go to Washington “with torches and pitchforks” and demand disclosure of what was tested and what the results were; he thinks this is unlikely to happen. The second is to evaluate predictions confidentially. Using its trial data, FDA could run AI models from different developers “behind a blind wall,” keeping the developers unaware of which trials were being used for evaluation. After accumulating thousands of predictions and waiting for the trials to finish, FDA would give developers aggregate feedback on their models’ performance.

A similar blinded evaluation has already been conducted: in the DREAM prostate cancer challenge, outcomes from one of four trials were withheld and used to evaluate 50 algorithms for predicting patient survival. Under FDA’s draft guidance, a company establishes the credibility of an AI model used to support its drug application; in Betts-LaCroix’s proposal, the agency would evaluate the models. Once models have made accurate predictions over a sustained period, he hopes regulators will say that AI has proposed a pill for an age-related disease and has been right in its previous predictions.

“Let’s just give this to people and no longer need a clinical trial. That’s the future I want.”

Retro itself is conducting a phase one trial of RTR242, a drug that restores autophagy (the process by which cells clear out their own waste); results could arrive around 2027, he said. A separate Retro program focuses on proteins designed by an AI model developed with OpenAI. The model’s first task was to improve the Yamanaka factors, proteins that return adult cells to a stem cell state, because Retro can reliably measure the outcome of this reprogramming. He said roughly half of the proteins generated by the model worked in cells, and some differed from their natural counterparts at up to 80% of amino acid positions. Asked by an audience member how these proteins work, Betts-LaCroix compared them to AlphaGo’s move 37 in its 2016 match against Go champion Lee Sedol: one commentator took the move for a mistake, but the program won the game.

“It’s something like move 37 in Go: there’s a new move that seems to win the game, but how exactly does it work?”